The Good, The Fast And The Better Pedestrian Detector

Abstract

Pedestrian detection is a well-known problem in Computer Vision, mostly because of its direct application in surveillance. In the past decade, several efforts have been performed to improve the detection in terms of accuracy, velocity and enhancement of features. In this work, we proposed and analyzed techniques focusing on these points. Firstly, we propose an oblique random forest associated with Partial Least Squares (PLS). The method consists on using the PLS to find a hyperplane at each decision tree node. Secondly, we evaluate filtering approaches to reduce the search space and keep only potential regions of interest in the scene. Finally, we propose an approach to combine distinct pedestrian detectors by reinforcing the human hypothesis whereas suppressing a significant number of false positives due to the lack of spatial consensus when multiple detectors are considered. Our proposed approach outperforms all previously published state-of-the-art methods.